Analytically guided machine learning for surface roughness determination in steel grinding processes

Abstract

This paper presents a machine learning (ML)-based approach for predicting surface roughness in various steel types during the grinding process. Three different algorithms—Gaussian process regression (GPR), gradient boosting (GB), and K-nearest neighbors (KNNs) regression—were applied, with a comprehensive set of processing parameters and analytical equations used as input features. The results show that the GPR model outperforms the other models, with an R² value of 0.964, compared to 0.938 for GB and 0.901 for KNN. Additionally, the GPR model achieved the lowest root mean squared error (RMSE) of 0.034, while GB and KNN had RMSE values of 0.058 and 0.096, respectively. Additionally, it is found that integrating analytical equations into the ML models significantly improved prediction accuracy by integrating domain knowledge and theoretical insights, which helped reduce bias and enhance model generalization. The study also revealed that the model's sensitivity to different factors varies with surface roughness. For smoother surfaces, the physical and mechanical properties of both the grinding tool and workpiece materials play a more significant role. In contrast, as surface roughness increases, the influence of geometric and processing parameters on prediction performance becomes more pronounced, emphasizing the importance of operational conditions in the grinding process. This research not only enhances the predictive capabilities of surface roughness models but also offers valuable insights into the factors influencing the grinding process.